Anthropic this week disclosed that it developed an internal version control system after daily code merges surged past 30,000 in September 2026, up from just 2,500 four months earlier in May. The AI firm revealed its solution, called Scion, in an announcement that detailed how Claude — its own AI coding assistant — was both responsible for the infrastructure bottleneck and essential to solving it. The company says it completed the entire build and migration of its monorepo to Scion in just five months, but won't be releasing the tool publicly anytime soon.
According to Anthropic, Scion has a 200 times smaller infrastructure footprint in CPU and memory compared to its previous git-mirroring setup. The new system also lets CI and research jobs retrieve code two orders of magnitude faster than before. The company attributes the explosion in code volume to Claude Mythos Preview and Claude Tag, both made available internally around the same time the Scion project began. A team of just seven people built Scion alongside a new code review tool named Antique and half a dozen other systems in those five months — work Anthropic says would typically require a large-scale engineering team years to complete.
The company wrote that "coding is no longer the bottleneck to building systems or doing research" at Anthropic, with Claude's ability to generate code at massive speed and scale now the reason developer infrastructure struggles to keep pace. The report states that "every bottleneck we remove immediately leads to a further increase in the amount of code we produce." Anthropic describes git as no longer suitable for high-volume AI coding, at least not for its enormous monorepo and exceptionally large volume of code changes. The company also claims Scion is "among the most well-tested and well-defended codebases we have ever worked on," crediting Claude's rapid coding capability with enabling extensive simulations, fuzzing, and sophisticated rate limiting.
Anthropic says Scion's design departs from git by providing always-on server access, which eliminates git's assumption that clients need offline work capability with complete repository copies — unnecessary overhead for clients that are always online. The system also offers API-based source control, allowing CI jobs or tools to query code history and other repository details without maintaining a local repository the way git does. Because Scion distributes repository data across independently scalable components, it can expand capacity as needed without maintaining hundreds of git mirrors. Despite predicting that other AI-heavy teams will soon face the same scaling challenges, Anthropic isn't making Scion publicly available and has committed to no public version, citing prohibitively high token costs and the tool's niche applicability. The company states that "Anthropic is unique in that our ability to spend tokens for internal use far exceeds what any other business would realistically want to spend," and calls Scion "by no means a suitable replacement for git for general use."
The company does believe other organizations will encounter similar infrastructure strain as AI coding agents grow more capable, noting that it's "often the first to see where tools built for human scale start to strain" because it uses Claude to write software more than almost any other company. Anthropic says it shared its approach to encourage developers to apply some of Scion's successful concepts to git, and expressed interest in partnering with GitHub, GitLab, and the open source community to bring some of those ideas to git. For now, inspiration is all Scion-curious developers can expect — the migration work required to adapt the system for other environments would be extensive, according to Anthropic. The implication for decision-makers is that infrastructure planning must now anticipate agent-generated workloads that dwarf human output, yet proprietary solutions may fragment the ecosystem rather than converge around shared standards.